6 perf-top - System profiling tool.
11 'perf top' [-e <EVENT> | --event=EVENT] [<options>]
15 This command generates and displays a performance counter profile in real time.
22 System-wide collection. (default)
26 Event period to sample.
30 Monitor only on the list of CPUs provided. Multiple CPUs can be provided as a
31 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
32 Default is to monitor all CPUS.
36 Number of seconds to delay between refreshes.
40 Select the PMU event. Selection can be a symbolic event name
41 (use 'perf list' to list all events) or a raw PMU event in the form
42 of rN where N is a hexadecimal value that represents the raw register
43 encoding with the layout of the event control registers as described
44 by entries in /sys/bus/event_source/devices/cpu/format/*.
48 Display this many functions.
51 --count-filter=<count>::
52 Only display functions with more events than this.
55 Sort the output by the event at the index n in group. If n is invalid,
56 sort by the first event. It can support multiple groups with different
57 amount of events. WARNING: This should be used on grouped events.
61 Profile at this frequency. Use 'max' to use the currently maximum
62 allowed frequency, i.e. the value in the kernel.perf_event_max_sample_rate
67 Child tasks do not inherit counters.
71 Path to vmlinux. Required for annotation functionality.
80 --mmap-pages=<pages>::
81 Number of mmap data pages (must be a power of two) or size
82 specification with appended unit character - B/K/M/G. The
83 size is rounded up to have nearest pages power of two value.
87 Profile events on existing Process ID (comma separated list).
91 Profile events on existing thread ID (comma separated list).
95 Record events in threads owned by uid. Name or number.
98 --realtime=<priority>::
99 Collect data with this RT SCHED_FIFO priority.
101 --sym-annotate=<symbol>::
102 Annotate this symbol.
105 --hide_kernel_symbols::
109 --hide_user_symbols::
113 Demangle kernel symbols.
117 Dump the symbol table used for profiling.
121 Be more verbose (show counter open errors, etc).
125 Zero history across display updates.
129 Sort by key(s): pid, comm, dso, symbol, parent, srcline, weight,
130 local_weight, abort, in_tx, transaction, overhead, sample, period.
131 Please see description of --sort in the perf-report man page.
134 Specify output field - multiple keys can be specified in CSV format.
135 Following fields are available:
136 overhead, overhead_sys, overhead_us, overhead_children, sample and period.
137 Also it can contain any sort key(s).
139 By default, every sort keys not specified in --field will be appended
144 Show a column with the number of samples.
146 --show-total-period::
147 Show a column with the sum of periods.
150 Only consider symbols in these dsos. This option will affect the
151 percentage of the overhead column. See --percentage for more info.
154 Only consider symbols in these comms. This option will affect the
155 percentage of the overhead column. See --percentage for more info.
158 Only consider these symbols. This option will affect the
159 percentage of the overhead column. See --percentage for more info.
162 --disassembler-style=:: Set disassembler style for objdump.
165 Path to addr2line binary.
168 Path to objdump binary.
172 Remove first N entries from source file path names in executables
173 and add PREFIX. This allows to display source code compiled on systems
174 with different file system layout.
177 Interleave source code with assembly code. Enabled by default,
178 disable with --no-source.
181 Show raw instruction encoding of assembly instructions.
184 Enables call-graph (stack chain/backtrace) recording.
186 --call-graph [mode,type,min[,limit],order[,key][,branch]]::
187 Setup and enable call-graph (stack chain/backtrace) recording,
188 implies -g. See `--call-graph` section in perf-record and
189 perf-report man pages for details.
192 Accumulate callchain of children to parent entry so that then can
193 show up in the output. The output will have a new "Children" column
194 and will be sorted on the data. It requires -g/--call-graph option
195 enabled. See the `overhead calculation' section for more details.
196 Enabled by default, disable with --no-children.
199 Set the stack depth limit when parsing the callchain, anything
200 beyond the specified depth will be ignored. This is a trade-off
201 between information loss and faster processing especially for
202 workloads that can have a very long callchain stack.
204 Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
206 --ignore-callees=<regex>::
207 Ignore callees of the function(s) matching the given regex.
208 This has the effect of collecting the callers of each such
209 function into one place in the call-graph tree.
212 Do not show entries which have an overhead under that percent.
216 Determine how to display the overhead percentage of filtered entries.
217 Filters can be applied by --comms, --dsos and/or --symbols options and
218 Zoom operations on the TUI (thread, dso, etc).
220 "relative" means it's relative to filtered entries only so that the
221 sum of shown entries will be always 100%. "absolute" means it retains
222 the original value before and after the filter is applied.
225 --column-widths=<width[,width...]>::
226 Force each column width to the provided list, for large terminal
227 readability. 0 means no limit (default behavior).
230 When processing pre-existing threads /proc/XXX/mmap, it may take
231 a long time, because the file may be huge. A time out is needed
233 This option sets the time out limit. The default value is 500 ms.
238 Enable taken branch stack sampling. Any type of taken branch may be sampled.
239 This is a shortcut for --branch-filter any. See --branch-filter for more infos.
243 Enable taken branch stack sampling. Each sample captures a series of consecutive
244 taken branches. The number of branches captured with each sample depends on the
245 underlying hardware, the type of branches of interest, and the executed code.
246 It is possible to select the types of branches captured by enabling filters.
247 For a full list of modifiers please see the perf record manpage.
249 The option requires at least one branch type among any, any_call, any_ret, ind_call, cond.
250 The privilege levels may be omitted, in which case, the privilege levels of the associated
251 event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege
252 levels are subject to permissions. When sampling on multiple events, branch stack sampling
253 is enabled for all the sampling events. The sampled branch type is the same for all events.
254 The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k
255 Note that this feature may not be available on all processors.
258 Add the addresses of sampled taken branches to the callstack.
259 This allows to examine the path the program took to each sample.
262 When displaying traceevent output, do not use print fmt or plugins.
266 Enable hierarchical output. In the hierarchy mode, each sort key groups
267 samples based on the criteria and then sub-divide it using the lower
270 For example, in normal output:
272 perf report -s dso,sym
274 # Overhead Shared Object Symbol
275 # ........ ................. ...........
276 50.00% [kernel.kallsyms] [k] kfunc1
278 15.00% [kernel.kallsyms] [k] kfunc2
280 5.00% libc.so [.] libcall
284 perf report -s dso,sym --hierarchy
286 # Overhead Shared Object / Symbol
287 # .......... ......................
288 65.00% [kernel.kallsyms]
298 Enable this to use just the most recent records, which helps in high core count
299 machines such as Knights Landing/Mill, but right now is disabled by default as
300 the pausing used in this technique is leading to loss of metadata events such
301 as PERF_RECORD_MMAP which makes 'perf top' unable to resolve samples, leading
302 to lots of unknown samples appearing on the UI. Enable this if you are in such
303 machines and profiling a workload that doesn't creates short lived threads and/or
304 doesn't uses many executable mmap operations. Work is being planed to solve
305 this situation, till then, this will remain disabled by default.
308 Don't do ownership validation.
310 --num-thread-synthesize::
311 The number of threads to run when synthesizing events for existing processes.
312 By default, the number of threads equals to the number of online CPUs.
315 Record events of type PERF_RECORD_NAMESPACES and display it with the
316 'cgroup_id' sort key.
320 monitor only in the container (cgroup) called "name". This option is available only
321 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
322 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
323 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
324 to first event, second cgroup to second event and so on. It is possible to provide
325 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
326 corresponding events, i.e., they always refer to events defined earlier on the command
327 line. If the user wants to track multiple events for a specific cgroup, the user can
328 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
331 Record events of type PERF_RECORD_CGROUP and display it with the
334 --switch-on EVENT_NAME::
335 Only consider events after this event is found.
339 Find out where broadcast packets are handled
341 perf probe -L icmp_rcv
343 Insert a probe there:
345 perf probe icmp_rcv:59
347 Start perf top and ask it to only consider the cycles events when a
348 broadcast packet arrives This will show a menu with two entries and
349 will start counting when a broadcast packet arrives:
351 perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
353 Alternatively one can ask for a group and then two overhead columns
354 will appear, the first for cycles and the second for the switch-on event.
356 perf top -e '{cycles,probe:icmp_rcv}' --switch-on=probe:icmp_rcv
358 This may be interesting to measure a workload only after some initialization
359 phase is over, i.e. insert a perf probe at that point and use the above
360 examples replacing probe:icmp_rcv with the just-after-init probe.
362 --switch-off EVENT_NAME::
363 Stop considering events after this event is found.
365 --show-on-off-events::
366 Show the --switch-on/off events too. This has no effect in 'perf top' now
367 but probably we'll make the default not to show the switch-on/off events
368 on the --group mode and if there is only one event besides the off/on ones,
369 go straight to the histogram browser, just like 'perf top' with no events
370 explicitly specified does.
373 Show callgraph with stitched LBRs, which may have more complete
374 callgraph. The option must be used with --call-graph lbr recording.
375 Disabled by default. In common cases with call stack overflows,
376 it can recreate better call stacks than the default lbr call stack
377 output. But this approach is not foolproof. There can be cases
378 where it creates incorrect call stacks from incorrect matches.
379 The known limitations include exception handing such as
380 setjmp/longjmp will have calls/returns not match.
383 --pfm-events events::
384 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
385 including support for event filters. For example '--pfm-events
386 inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
387 option using the comma separator. Hardware events and generic hardware
388 events cannot be mixed together. The latter must be used with the -e
389 option. The -e option and this one can be mixed and matched. Events
390 can be grouped using the {} notation.
393 INTERACTIVE PROMPTING KEYS
394 --------------------------
397 Display refresh delay.
400 Number of entries to display.
403 Event to display when multiple counters are active.
406 Profile display filter (>= hit count).
409 Annotation display filter (>= % of total).
415 Stop annotation, return to full profile display.
424 Toggle event count zeroing across display updates.
429 Pressing any unmapped key displays a menu, and prompts for input.
431 include::callchain-overhead-calculation.txt[]
435 linkperf:perf-stat[1], linkperf:perf-list[1], linkperf:perf-report[1]